Tag
Introduces TASER, a training-time regularization framework derived from Langevin Stein operators that encourages geometric compatibility between predictors and data density, improving adversarial robustness and stability on CIFAR-10 without significant clean accuracy degradation.